refresh

Refresh Skill

Incremental update of an existing analysis. The server already persists prior state; this skill computes the delta, re-classifies only stale files, and re-synthesizes.

Prerequisites

A prior analysis must exist for this project. If none does, call the analyze skill instead — there's nothing to refresh from a blank slate.

Pipeline

Step 1: Read prior analysis

Call read_analysis with the repo path. The response contains:

• json — the full prior analysis (or null if none).
• markdown — file-classifier-format rendering of the prior file analysis.
• freshness — {staleFiles, missingFiles, removedFiles, staleSynthesis, computedAtHead, computedAtBase}.

If json is null, invoke the analyze skill instead — there's nothing to refresh.

If freshness.staleFiles, missingFiles, and removedFiles are all empty, report "analysis is fresh, nothing to do" and exit.

Step 2: Write prior markdown to scratch

Write the markdown field from read_analysis to /tmp/lgtm-analysis-files-prev.md. This becomes the agent's prior-context input.

Step 3: Spawn file-classifier in delta mode

Spawn the file-classifier agent with a prompt of the form:

Refresh analysis for the repository at <REPO_PATH>.
The base branch is <BASE_BRANCH>.

Previous analysis (for category/style continuity):
/tmp/lgtm-analysis-files-prev.md

Re-classify ONLY these files (diffs have changed since last analysis):
- <each path from staleFiles ∪ missingFiles, one per line>

Read the agent's documented "Delta mode" section in its instructions: it should
output ONLY entries for the listed files to /tmp/lgtm-analysis-files.md. Do NOT
copy unchanged entries from the previous analysis — the server merges them in.

Find the base branch the same way the analyze skill does:

1. Try gh pr view --json baseRefName -q .baseRefName. If it succeeds, fetch with git fetch origin <branch> and use origin/<branch>.
2. Fall back to main (or master) if there's no open PR.

Step 4: Compose merged file analysis

Locally compose the merged file analysis (previous entries minus removedFiles, overlaid with the new entries the agent wrote) and save to /tmp/lgtm-analysis-files-merged.md. This is the input to the synthesizer.

A simple recipe: take /tmp/lgtm-analysis-files-prev.md, drop any ## <path> block where <path> appears in removedFiles ∪ staleFiles ∪ missingFiles, then append the entire content of /tmp/lgtm-analysis-files.md.

(You can do this with Read + string slicing in a small bash command, or read both files via the Read tool and paste them together with the Write tool.)

Step 5: Spawn synthesizer

Same as in /lgtm analyze:

Read the file analysis at /tmp/lgtm-analysis-files-merged.md.
Session description: <DESCRIPTION or "No description provided.">
Write your synthesis to /tmp/lgtm-analysis-synthesis.md
Write your review guide to /tmp/lgtm-analysis-review-guide.md

The synthesizer agent is unchanged — re-uses the existing pipeline.

Step 6: Submit

Call set_analysis with:

• repoPath: the repo path
• fileAnalysisPath: /tmp/lgtm-analysis-files.md (the agent's delta-only output)
• synthesisPath: /tmp/lgtm-analysis-synthesis.md
• reviewGuidePath: /tmp/lgtm-analysis-review-guide.md
• mode: "merge"
• removedFiles: freshness.removedFiles

The server merges, stamps blob SHAs, and broadcasts analysis_changed so connected browsers refresh.

Reporting

Tell the user:

• N files re-classified (|staleFiles ∪ missingFiles|)
• M files dropped (|removedFiles|)
• Total file count after merge
• Synthesis re-run

If set_analysis returns an error, read the relevant markdown file to diagnose, fix the agent's output, and retry.